Dual-Population Hybrid Genetic Programming for Decision Forest Optimization in Adaptive Classification Rule Construction

Abstract

Complex data fusion requires identifying consistent associations across heterogeneous data sources with diverse semantics, structures and feature representations, posing significant challenges for reliable classification. Decision Tree (DT)-based methods provide interpretability but often lack robustness and adaptability in highly heterogeneous environments, while ensemble approaches such as Decision Forests (DFs) still face challenges in balancing diversity, efficiency and optimization complexity. This paper aims to develop an adaptive classification framework for complex data fusion by enhancing structural diversity and evolutionary optimization. To this end, a Dual-Population-based Hybrid Genetic Programming (DPHGP) framework is proposed to automatically construct and optimize DF-based classification rules. Specifically, a dual-population co-evolutionary mechanism is designed to evolve decision trees for similarity feature construction and ensemble structures for aggregation, thereby improving diversity and robustness. Meanwhile, adaptive breeding operators are introduced to balance global exploration and local exploitation during evolution. Furthermore, a Compact Genetic Algorithm (CGA) is incorporated to refine numerical constants within elite classifiers, thereby supporting more effective numerical parameter optimization. The experimental results on OAEI benchmark datasets show that DPHGP achieves competitive matching accuracy and robustness across the evaluated datasets, compared with state-of-the-art evolutionary and machine learning approaches. These results indicate that the proposed framework provides an effective and interpretable solution for adaptive rule construction in complex data fusion scenarios.